Comment by PeterisP

For some reason they focus on the inference, which is the computationally cheap part. If you're working on ML (as opposed to deploying someone else's ML) then almost all of your workload is training, not inference.